Documente Academic
Documente Profesional
Documente Cultură
I. INTRODUCTION-I
Since analog-to-digital converters (ADCs) ultimately limit the Fig. 1. Time-interleaved ADC (TI-ADC) with M channels. Each channel
performance of today's communication systems, high-speed, high- alternately takes samples at a rate M from the input signal xa (t). At the
resolution, and power-aware ADCs are required in order to comply multiplexer (MUX), the samples from the M parallel channels are merged
with new communication standards. This also leads to an increased into one output channel running at an M times higher rate fs.
demand for high-speed and high-resolution sampling systems in the
measurement industry [1]. Present ADC technologies work on their
limits and cannot be properly pushed further, since the downscaling and the multiplexer (MUX) to recombine the digital outputs of the
of IC technologies to deep sub-micron technologies makes their channels. The conversion rate of the overall system is increased by
design even more difficult. However, the increased component density the number of channels M. It should be noticed that each channel
of digital circuits allows for using additional chip area with small has to deal with the entire input signal xa (t), and, therefore, the
additional costs [2]. sample-and-holds in each channel have to resolve the full input signal
One possibility to overcome these performance limits is to use bandwidth.
parallelism, i.e., to split the information of the analog input signal From a theoretical point of view, we can increase the sampling
into several parallel channels, to convert them independently, and rate of a TI-ADC by the number of channels that work in parallel in
finally to recombine them into one digital output signal. In theory, the system. Ideally, the sampling rate would linearly scale with the
which was introduced by Papoulis' Generalized Sampling Expansion number of channels; however, channel mismatches ultimately limit
(GSE) [3], there are many ways to split the information of the the performance of TI-ADCs. On the one hand, the downscaling of
input signal. In practice, only a few parallel multi-channel sampling the IC technologies complicates the matching of the components,
structures [4] have been further analyzed [5]-[7], where the time- but, on the other hand, the increased component density allows for
interleaved structure is among the most promising ones for the future. including additional digital components with small additional costs.
The idea of a time-interleaved ADC (TI-ADC) is that each channel Therefore, we can add digital circuits to overcome the problems of
in a system of M parallel channels alternately takes one sample, analog converter circuits [11]. TI-ADCs constitute a prime example
whereas the sampling frequency of one channel does not need to of such merging technologies, where the technology can only be
fulfill the Nyquist Criterion [8]. However, when in the digital domain properly pushed further, when we consider digitally enhanced analog
all samples merge into one sequence we obtain an overall sampling circuits.
frequency that fulfills the Nyquist criterion. Thus, sampling with an
ideal TI-ADC with M channels is equivalent to sampling with an II. CHANNEL MISMATCHES
ideal ADC with an M times higher sampling rate. The channels Each channel ADC in a TI-ADC has technology dependent errors
of a TI-ADC can be realized in different converter technologies to (e.g., integral nonlinearity errors, clock jitter) like a single-channel
achieve for example high-rate and low-power ADCs [9] or high-rate ADC, but due to component mismatches among the channels, ad-
and high-resolution ADCs [10]. ditional errors, called mismatch errors, are introduced [12]. This is
The typical structure of a TI-ADC is shown in Fig. 1. We see the illustrated in Fig. 2, where we see a TI-ADC with channel mismatches
analog input signal Xa (t), the M time-interleaved parallel channels, and without channel mismatches for a sinusoidal input signal. For
MT Ys>JQ)T2~
=kK E V (Q- J M ) XaK(iQ-ikM,)J' (3)
| ~ ~~~
too M1/ T, iT, yo(t)
<
lwhere1M-
Vk (iQ) = M E Hm (jQ) em M, (4)
mT, yo,ev(t)
Q Without an input signal the sampled output, i.e., offset mismatches
xt,,(t) IMT E y,(t)nmy, only, is
M ~
+s~(M1T
[+t I |M- n=o
60 whichareweightedwhere
Pulse atthek Fourier transform of (5) gives
3387
Signal with Signal without
channel mismatches channel mismatches can tune the matching on the analog side or we can reconstruct the
distorted signal on the digital side. It is also possible to combine both
approaches [21].
Digital calibration is attractive in many ways. The digital cali-
Tune channel Adapt bration is, like the principal topology of a TI-ADC, independent
xj(t) My[
properties
TI-ADC
j
Normazefrequen]y Iparameters
Digital algorithm
Nomazefjequ.n].
y from the used channel converter technology. Hence, we can apply
the same digital calibration method to TI-ADCs with different chan-
I f nel converter technologies, since the TI-ADC environment and the
--
l______ mismatch
functionality of the
production pprocess do not ddirectly influence the functionality
L calibration method.
-
identification
analog digital A. Correction Methods
Fig. 5. Possibilities to calibrate a TI-ADC. Although the correction of gain and offset mismatches in the
digital domain is quite simple, since we only have to add at most
one adder and one multiplier to the signal path of each channel
Thus, the gain mismatch is the deviation of the gains gm from the ADC, the correction of timing mismatches (linear-phase mismatches)
average gain of all channels. In practical TI-ADC designs, however, is much more difficult. In fact, it is a sub-problem of the non-
we have to deal with magnitude mismatches. To compensate for them, uniform sampling problem. For TI-ADCs the problem simplifies to
a first solution is to use some kind of average magnitude response periodically non-uniform sampled signals, i.e., the time shifts Atm
for each channel over the frequency band of interest instead of DC exhibit a periodicity, where the time shifts Atm are small compared
gains. to the sampling period T8. However, under the constraint of an on-
The timing mismatch is the deviation from the averaged linear- chip implementation the problem becomes difficult again.
phase responses of the channels normalized by the frequency. To see For the timing-mismatch problem accurate solutions have been
this, we split the phase responses into a linear and a nonlinear part, found in [19], [20], [22], [23], although only for some of them [19],
i.e., [20] the implementation on a TI-ADC chip is maintainable. However,
(m (Q) = tmQ + >m (Q)m (10) for changing timing mismatches the used reconstruction method has
to be easily adaptable. Thus, an open question is to find reconstruction
where tmQ is the linear-phase response over the frequency band of methods where the needed coefficients can be simply derived from the
interest. The timing mismatch Atm is the deviation from the averaged estimated timing mismatches. A first solution to this problem can be
and frequency normalized linear-phase response, i.e., found in [21], [24]. In [24] the authors show that by using fractional
1 M-1 delay filters they only have to redesign one coefficient in each channel
Atm = tm M E tm. (11) to adapt to changed timing mismatches and in [21] a method was
mM=0 introduced which reorders the channel sequence in order to achieve
It should be noticed that we can treat aperture-delay mismatches with a spectrally shaped output signal. Unfortunately, both methods need
this model as well. The aperture-delay mismatch is the deviation from some amount of additional oversampling.
the ideal sampling instant caused by time-shifted clock signals. This Therefore, the goal of timing mismatch correction is to find an ac-
delay can be represented by an equivalent time-shift (linear phase curate, power-aware method, which only needs a slight oversampling
shift) of the input signal in each channel which can be accomplished and which can be easily adapted to changing timing mismatches.
by the linear filters Hm (jQ). If these problems are solved, magnitude and nonlinear-phase mis-
matches (bandwidth mismatches) will limit the effective resolution
III. CALIBRATION OF CHANNEL MISMATCHES of TI-ADCs and will therefore have to be corrected for a further
Avoiding mismatches is the main concern in designing fast TI- improvement [18].
ADCs [9]. Unfortunately, shrinking IC technologies and increasing B Identification M
clock rates make component matching even more difficult. Further- ethods
more, the matching is influenced by time-variant parameters such as The identification of mismatch parameters is the most critical
temperature or component aging. Therefore, calibration methods for component in the channel mismatch compensation process of TI-
TI-ADCs have been proposed, which tune the component matching, ADCs. If the identified parameters are wrong even the best correction
e.g., [14], [15], or digitally correct the distorted output signal, method cannot improve the TI-ADC performance.
e.g., [16]-[20]. For the identification of the channel mismatches with special input
In Fig. 5 the principal calibration methods are illustrated. We see signals we can find accurate solutions [18], [25], [26]. Nevertheless,
a TI-ADC driven by a sinusoidal input signal xa (t). At the output identification with special input signals is suitable for measurement
of the TI-ADC (y[n]) we see the sampled spectrum of the sinusoidal applications with calibration cycles but not suitable for communi-
input signal and distortions caused by channel mismatches. In order to cations systems, where in general we have no extra duty cycles
reduce these distortions we have to identify the significant mismatch for the calibration. There, the identification has to be done during
parameters. This task is much easier to realize when we know the the normal operation of the TI-ADC with the only restriction of
input signal Xa (t), which is indicated by a dashed arrow from the bandlimited input signals. For single communication protocols, we
identification box to the input signal. In many cases, however, we can assume particular signal statistics, however, the trend towards
have no or just little statistical knowledge, e.g., bandlimitation or software defined radio (SDR) does not allow such assumptions.
modulation technique, about the input signal. This little knowledge For gain and offset mismatch identification we can find some
makes it much more difficult to obtain reliable estimates for the methods [17], [27], [28]. In many cases, we obtain good offset
mismatch parameters. If the mismatch parameters are known, we and gain mismatch estimates with a simple comparison of the
3388
averaged output values and the averaged output power among all [11] A. van Roermund, H. Hegt, P. Harpe, G. Radulov, A. Zanikopoulos,
channels. Unfortunately, such methods are vulnerable to an input K. Doris, and P. Quinn, "Smart AD and DA converters," in IEEE
signal correlation with the switching sequence of the channels, i.e., International
pp. 4062-4065.
Symposium on Circuits and Systems, vol. 6, May 2005,
we onlywegetolya gt a imitdnuber
limited number off dffernt
different smple
samples fro eac chanel
from each channel [12] C. Vogel, "The impact of combined channel mismatch effects in time-
for the estimation process. interleaved ADCs," IEEE Trans. on Instrumentation and Measurement,
The most challenging problem is the on-line identification of vol. 54, no. 1, pp. 415-427, February 2005.
timing mismatches. The identification should be accurate and its [13] C. Vogel and G. Kubin, "Modeling of time-interleaved ADCs with
nonlinear hybrid filter banks," AEU-International J. of Electronics and
implementation should be cost-efficient. Although we can find identi- Communications, vol. 59, no. 5, pp. 288-296, 2005.
fication methods in the literature [17], [29], they are either imprecise, [14] K. Dyer, F. Daihong, S. Lewis, and P. Hurst, "An analog background
limited in the number of channels, or have an enormous computa- calibration technique for time-interleaved analog-to-digital converters,"
tional complexity. IEEE J. of Solid-State Circuits, vol. 33, no. 12, pp. 1912-1919, Decem-
ber 1998.
Hence,Hence
we weneedeedstable
sableon-lie
on-line idntifcatio metods fr
identification methods ofset,
for offset, [ 15] C. Vogel, D. Draxelmayr, and F Kuttner, "Compensation of timing mis-
gain, and timing mismatches, where an accurate and implementation matches in time-interleaved analog-to-digital converters through transfer
efficient timing mismatch identification is the foremost challenge for characteristics tuning," in Proceedings of the 47th IEEE International
the future. After having solved these problems, however, we will Midwest Symposium On Circuits and Systems, vol. 1, July 2004, pp.
also have to deal with the identification of bandwidth mismatches 341-344.
[16] H. Jin and E. K. F. Lee, "A digital-background calibration technique
and nonlinearity mismatches. for minimizing timing-error effects in time-interleaved ADCs," IEEE
Trans. on Circuits and Systems II: Analog and Digital Signal Processing,
IV. CONCLUSION vol. 47, no. 7, pp. 603-613, July 2000.
[17] S. Jamal, D. Fu, M. Singh, P. Hurst, and S. Lewis, "Calibration of
We have presented a unified framework for dealing with linear sample-time error in a two-channel time-interleaved analog-to-digital
channel mismatches and offset mismatches. In particular, we can converter," IEEE Trans. on Circuits and Systems I: Regular Papers,
use this framework to describe gain and timing mismatches as a vol. 51, no. 1, pp. 130-139, January 2004.
[18] M. Seo, M. Rodwell, and U. Madhow, "Comprehensive digital correction
of mismatch errors for a 400-Msamples/s 80-dB SFDR time-interleaved
and identification of channel mismatches and we have pointed out analog-to-digital converter," IEEE Trans. on Microwave Theory and
the most challenging problems for the near future. Since the major Techniques, vol. 53, no. 3, pp. 1072-1082, April 2005.
problems of offset, gain, and timing mismatch correction have been [19] H. Johansson and P. Lowenborg, "Reconstruction of nonuniformly
recently solved, the main concern for the future will be finding stable sampled bandlimited signals by means of time-varying discrete-time
FIR filters," to appear in J. Applied Signal Processing - Special Issue
and reliable on-line mismatch identification methods. However, to on Frames and Overcomplete Representations in Signal Processing,
achieve very high-resolution TI-ADCs more errors, e.g., bandwidth Communications, and Information Theory, 2006.
mismatches, INL mismatches, and DNL mismatches, have to be [20] R. Prendergast, B. Levy, and P. Hurst, "Reconstruction of band-limited
considered and calibrated as well. For example, for a 100 MHz periodic nonuniformly sampled signals through multirate filter banks,"
bandwidth, a50MH DCbadiIEEE
Trans. on Circuits and Systems I: Regular Papers, vol. 51, no. 8,
signal bandwidth, a 750 MHz ADC bandwidth, andsignal an analog device pp. 1612-1622, August 2004.
matching of 1%, the bandwidth mismatch will limit the resolution to [21] C. Vogel, D. Draxelmayr, and G. Kubin, "Spectral shaping of timing mis-
some 12-13 bits. The research on these mismatches has just started. matches in time-interleaved analog-to-digital converters," in Proceedings
of the 2005 IEEE International Symposium on Circuits and Systems, May
REFERENCES 2005, pp. 1394-1397.
[22] Y C. Jenq, "Perfect reconstruction of digital spectrum from nonuni-
[1] K. Poulton, R. Neff, B. Setterberg, B. Wuppermann, T. Kopley, R. Jewett, formly sampled signals," IEEE Trans. on Instrumentation and Measure-
J. Pernillo, C. Tan, and A. Montijo, "A 20 GS/s 8 b ADC with a 1 MB ment, vol. 46, no. 7, pp. 649 - 652, June 1997.
memory in 0.18 ,um CMOS," in IEEE International Solid-State Circuits [23] Y C. Eldar and A. V. Oppenheim, "Filterbank reconstruction of bandlim-
Conference, vol. 1, February 2003, pp. 318-496. ited signals from nonuniform and generalized samples," IEEE Trans. on
[2] B. Murmann and B. E. Boser, Digitally Assisted Pipeline ADCs: Theory Signal Processing, vol. 48, no. 10, pp. 2864-2875, October 2000.
and Implementation. Springer, 2004. [24] H. Johansson and P. Lowenborg, "Reconstruction of nonuniformly
[3] An
3 A. Papoulis, "Generalized
"n z ssampling eexpansion," IIEEE Trans.s. onX Circuits
sampled bandlimited signals by of digital fractional delay filters,"
means
3389